Seph's Illustrious - Cinematic Anime
Preview images are "full final", only prompts (no LoRA's) -> ADetailer & Upscaled!
If you feel that there are too few preview images, enable rating R, as I usualy aim for that.
Aims for:
dark with vibrant colors (goth, neo-/noir)
generaly artistic images
2.5d characters (realistic, but still anime)
Settings:
For my full workflow (with Forge and DynamicPrompts), please see: https://civarchive.com/articles/14013
Sampler: Euler & Simple
CFG: 5-8
Steps: 12-14 as draft/preview, 24, 32, 42 for finals, 64 for super favorites
Aspect Ratio: should work with most, preview images are: 9:16 , 21:9 / render at: 720x1280, 1536x640, or 720x1536 (10:16 -> cellphone "portrait" wallpaper)
Prompt:
+ high_quality, highres, beautiful, detailed, masterpiece, best_quality, absurdres, 8k, HDR , very aesthetic , intricate_details , detailed_hands, elegant_fingers
- poorly_detailed, jpeg_artifacts, worst_quality, bad_quality
Happy imagination!
Also, I really would like to see what you guys created! :)
History / Recommendation:
All statements are in regards to 21:9.
Meaning either 1536x640 or 1720x720 "draft" renders (14 steps)!
v1:
Initial release
v2-experimental:
Raw quality droped - was workaround-able with img2img upscale + adetailer
Tried for slighly more realism, worked
Tried to preserve anime char, failed
Tried to improve "adult appearance" (on certain characters), failed
Failed to preserve the "intense black/colors"
v3-VAE:
Raw quality improved (back as v1 or even better)
Preserved "intense black/colors", yes
Age up characters, success (for most ; others are at least better than they were in v2)
Prompt adherence better than v2
Recommended IF: you prefer high quality characters with decent but obviously AI landscape/backgrounds
v4-VAE-fp16
Better "dark/black"
with few exceptions, most character appear more adult (at least when prompted for)
better backgrounds
either good "prompt adherence" or "good misunderstanding"
Info: stay at CFG 6 or lower, seems by increasing contrast an "overcooking" has happend, allthough effects are rare.. about 1-2% at CFG:6-8.. HOWEVER, for local gen, this chance can be totaly worth it as it also produces incredible images.-->EDIT: This was caused because i prompted for things, that are no longer required (as much).Recommended IF: you prefer characters with high quality landscape/backgrounds, while "full body" is not as good, portrait/macro shots are awesome
Description
Raw quality improved (back as v1 or even better)
Preserved "intense black/colors", yes
Age up characters, success (for most ; others are at least better than they were in v2)
Prompt adherence better than v2
FAQ
Comments (7)
THE BeST
THE most complete checkpoint I've seen. Not only fantastic detail, but varied and nuanced expressions, good accurate shadows, great depth of field, great material/hair textures, great overall lighting, great spot lighting, vibrant color, sharp focus (where intended), the fog/smoke effect was amazing, handles portrait and landscape well, and does well with scenery. Great job and great job showcasing the capabilities!
I haven't used it yet but, from the large amount of showcased images, I'd suggest that those testing it to start with low weights on whatever loras you use and just let it shine. And unless they're for something very specific, try it without loras and just use prompting. You may be quite pleasantly surprised at what this checkpoint can do. If you're skilled at balancing loras though, you may find that this check helps take your designs to a new level.
I'd be happy to "hear" some more feedback once you've used it ;)
This said, all my images shown here did NOT use any LoRA, but ONLY prompting!
Keep in mind, for most LoRA's one should use a weight of about 0.7 - 0.8
@Sephrajin It could be a while as I have low VRAM and some other system complications, but I'm looking forward to trying it out!
@HMQart28 My computer is almost 10 years old (i7-6700k, 16gb ram), and i'm using a 3060 Ti with 8GB VRAM, so I as well work on the bare minimum.
Either way, good luck!
@Sephrajin Interesting! Thank you for letting me know! How long does a local gen take? And do you train models locally or through a site? Also, are you mainly using comfy, a1111, or forge?
@HMQart28 i'm using forge, and 3 batches of 2 takes about 3 minutes for those 6 draft images at 1536x640.
For the full process, i've written that article that is linked in the description.